Quantifying Uncertainty in Dense Neural Networks Using Monte Carlo Dropout in Weather Predictions

Accurate and reliable weather forecasts are crucial for decision-making in sectors such as agriculture, disaster management, and transportation. Deep learning models have shown their promise in weather prediction tasks. However, since these models are used for decision-making in weather forecasting, they require more confidence in their predictions. Using uncertainty quantification methods for deep neural networks can increase the reliability of their predictions. This study uses Monte Carlo Dropout to estimate uncertainty for weather predictions, especially sunshine hours. To do so, a dense neural network model incorporating Monte Carlo Dropout was trained to forecast sunshine hours using historical data from 18 European cities from 2000 to 2010. Multiple forward passes using various dropout configurations produce a prediction distribution of predicted sunshine hours. The mean of this distribution is used as the final prediction, while the standard deviation quantifies the uncertainty of the model. We compared the performance of our proposed model with a traditional convolutional neural network. Experimental results indicate that the proposed model is robust to overfitting, exhibits stable learning behaviors, and improves predictive reliability with a standard deviation of 0.2469–2.529 compared to traditional convolutional neural networks. Evaluation with root mean squared error and mean absolute error confirmed that the proposed model outperformed the traditional convolutional network in predictive accuracy. The findings of this study highlight the importance of incorporating uncertainty quantification in meteorological models to support better decision-making and climate adaptation strategies.